Skip to content
Reference Guide

How Aura rewrites text, and where the model actually runs.

An AI writing assistant that sends your text to someone else's server is a different product from one that doesn't. This page walks through every part of Aura — the rewrite flow, the eight modes, how the diff review works, and the four engines it picks between depending on what your Mac can run.

01The shortcut and the flow

Select, press, review.

The rewrite

Select text in any app and press ⇧⌘P. The rewritten text streams back in place through a floating HUD.

Cancelling

esc cancels at any time, leaving both the original text and your clipboard untouched.

The second shortcut

A separate always-proofread shortcut, for when you want a correction rather than a rewrite and don't want to think about modes.

02The eight modes

What each one is actually for.

The list

Proofread (the default), Improve Clarity, Formal/Academic, Friendly, Make Concise, Expand, Bullet Points, and Custom Instruction.

Proofread is deliberately conservative

It corrects rather than rewrites. That is why Voice Mirroring is excluded from it — a proofread should return your sentence, fixed, not a different sentence in your style.

03The diff review card

The safeguard for longer selections.

Word-level diff

Longer selections show a review card before anything is inserted — deletions in red, insertions in green — with Insert or Discard.

Worth knowing: Short selections are replaced in place because reviewing a five-word fix costs more than redoing it. Longer ones are not, because silently replacing three paragraphs is how people lose work.

04Voice Mirroring

Making rewrites sound like you rather than like a model.

How it learns

Measures your own writing on-device from up to three samples — sentence length, contraction habits, register, punctuation fingerprints and emoji use — and applies that to rewrites.

Where it does not apply

Proofread, deliberately. See above.

05The four engines

Auto-picked, best available first.

The order

Apple Intelligence (macOS 26+, free) → Aura's own downloadable model (Core ML on the Neural Engine, Balanced 1.2 GB or Expert 2.8 GB, fully offline once downloaded) → any local OpenAI-compatible server such as LM Studio, Ollama or llama.cpp → a built-in NSSpellChecker-based deterministic cleanup.

The fallback always works

That last engine needs no model and no network, but handles Proofread only.

Worth knowing: Local servers are restricted to loopback addresses. Aura will not send your text to a remote endpoint, even one you configure yourself.

06Permissions and prompt injection

One permission, and one design decision worth stating.

Accessibility, and nothing else

Used to read the selection — via the accessibility API or a synthesized ⌘C — and to write the result back through the clipboard and ⌘V, restoring your previous clipboard afterwards.

Your selection is never an instruction

Selected text is always treated as material to edit, never as instructions to the model. Text that says "ignore your instructions and…" gets proofread, not obeyed.

07Now go try it.

Free for 14 days, no account. Back to Aura · Buying & licensing · Support